1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-24 03:33:20 +01:00

X86: add patterns for X86ISD::VSHLV and X86ISD::VSRLV

Replace VSELECT instruction which zeroes their result on exceeding legal SHL/SRL shift amount.
This commit is contained in:
Nekotekina 2018-07-06 00:24:31 +03:00
parent bed700114e
commit d5bc359dfd

View File

@ -42254,6 +42254,38 @@ static SDValue combineSelect(SDNode *N, SelectionDAG &DAG,
}
}
// Detect pattern for AVX2+ variable shifts (shl, lshr) for inf precision.
if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
SupportedVectorVarShift(VT.getSimpleVT(), Subtarget, ISD::SHL)) {
ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
// Check if one of the arms of the VSELECT is a zero vector. If it's on the
// left side invert the predicate to simplify logic below.
SDValue Other;
if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
Other = RHS;
CC = ISD::getSetCCInverse(CC, VT.getVectorElementType());
} else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
Other = LHS;
}
// Look for the following patterns (>> becomes vsrlv):
// y < 32 ? x << y : 0 --> vshlv(x, y)
// y <= 31 ? x << y : 0 --> vshlv(x, y)
APInt CondRHS;
if (Other && Other.getNumOperands() == 2 &&
DAG.isEqualTo(Other.getOperand(1), Cond.getOperand(0)) &&
(Other.getOpcode() == ISD::SHL || Other.getOpcode() == ISD::SRL) &&
ISD::isConstantSplatVector(Cond.getOperand(1).getNode(), CondRHS)) {
// Replace ISD::SHL or ISD::SHR with appropriate AVX2 vector-vector shift.
unsigned op = Other.getOpcode() == ISD::SHL ? X86ISD::VSHLV : X86ISD::VSRLV;
if ((CC == ISD::SETULT && CondRHS == VT.getScalarSizeInBits()) ||
(CC == ISD::SETULE && CondRHS == VT.getScalarSizeInBits() - 1))
return DAG.getNode(op, DL, VT, Other.getOperand(0), Other.getOperand(1));
}
}
if (SDValue V = combineVSelectWithAllOnesOrZeros(N, DAG, DCI, Subtarget))
return V;